An AI risk assessment is a structured review of every AI system your organisation builds, buys or uses, scored against the harm it could cause and the controls you have in place to stop that harm. It produces an inventory, a risk rating for each system, a gap analysis against a recognised standard and a mitigation plan with named owners and dates.
Businesses need one because in many organisations, AI has arrived by stealth. It came in through tools staff signed up for on a company card, AI features in a third-party software release or a marketing agency adding a chatbot to the website. It did not come through a controlled procurement process that asked what data the model sees, where that data goes, or who is accountable when the output is wrong. The distance between what a business thinks it is running and what it is actually running is the whole problem, and the regulatory exposure follows from it.
AI is already inside your business, whether or not you approved it
The Office for National Statistics recently published its combined view of AI in UK businesses covering 2023 to 2026. Adoption among businesses with 10 or more employees has close to tripled since September 2023, while depth has barely moved: the average adopting business used around 1.4 AI technologies in 2023 and around 1.6 in June 2026. Large language models lead at 18% of businesses, followed by visual content creation at 16%, machine learning for data processing at 12% and image processing at 6%.
Those figures count what businesses know about and are willing to report, but there’s a problem.
The figures miss the note-taker that joins client calls and stores transcripts on a server nobody has asked about, the CV ranking a recruitment platform added to its standard plan, and the contract someone pasted into a free chatbot at eight in the evening to get a summary before a meeting.
This is shadow AI, and it is a common issue in a first assessment.
The rules that apply to a London business in 2026
There is no UK AI Act. Existing regulators handle AI within their remits, so the obligations on a London business are somewhat scattered. The Department for Science, Innovation and Technology published its Blueprint for AI regulation on 21 October 2025, centred on the AI Growth Lab and its sector-by-sector regulatory sandboxes. Sandboxes help innovators test new products but are not a compliance route for AI you already run in production.
The Information Commissioner’s Office is where most of the practical pressure comes from. It published its AI and biometrics strategy, “Preventing harm, promoting trust”, in June 2025. From 12 May 2026, SI 2026/425 places the Commissioner under a statutory duty to produce a single code of practice on artificial intelligence and automated decision-making. The Data (Use and Access) Act 2025 rewrote the automated decision-making rules, and the ICO has been consulting on guidance for the new UK GDPR Articles 22A to 22D. It has also audited AI recruitment tools and found that some providers did not test adequately for accuracy and bias.
The EU AI Act reaches London businesses that place AI systems on the EU market or whose output is used in the EU.
A recruitment firm in Farringdon screening candidates for a Dublin office is likely to be in scope, as is an insurer underwriting EU risks and a SaaS company with customers in Berlin. The precise position depends on the system, how it is used and whether the business is acting as a provider or deployer.
The Act came into force on 1 August 2024, and the Digital Omnibus on AI, adopted by Parliament on 16 June 2026 and the Council on 29 June 2026, moved several dates. The timetable now runs:
- 2 August 2026: Article 50 transparency obligations apply, covering chatbot disclosure, deepfake labelling and marking of AI-generated content.
- 2 December 2026: Article 50(2) marking requirements extend to systems already on the market, and the new prohibited practices apply.
- 2 August 2027: member states must have at least one national AI regulatory sandbox in place.
- 2 December 2027: high-risk obligations apply to stand-alone Annex III systems, covering employment, education, credit scoring, essential services and critical infrastructure.
- 2 August 2028: high-risk obligations apply to AI embedded in the regulated products listed in Annex I.
Read the 2027 date carefully. The Omnibus extended the timeframe for Annex III high-risk systems by 16 months and changed none of the underlying obligations. Conformity assessment, technical documentation, data governance evidence and human oversight design all still have to be built, and firms that treat the extension as a pause will be doing that work in a hurry in late 2027.
Regulated firms in the City carry a further layer. The Prudential Regulation Authority’s supervisory statement SS1/23 on model risk management principles for banks took effect in May 2024 and brings AI and machine learning models within scope. Model risk expectations in financial services predate any AI-specific law, and supervisors can ask how a generative model in a client-facing process was validated.
What actually goes wrong
Most AI incidents are hidden. They run for months and are found by a customer rather than by a control.
Bias in decisioning has the clearest legal edge. A model trained on ten years of hiring data reproduces ten years of hiring patterns, and if those patterns disadvantaged a protected group, the model will too, consistently and at speed. The Equality Act 2010 does not care that a vendor supplied the algorithm.
Hallucination matters wherever output reaches a customer without a human reading it first. A support chatbot that invents a refund policy may expose the business to complaints, consumer-law questions and arguments over what the customer was promised, while a tool that invents a clause in a client letter turns a technical fault into a potential professional negligence question.
Prompt injection has no traditional equivalent. Instructions hidden in a document, an email or a web page can redirect a model that has been given access to systems and data. An assistant with mailbox access and no instruction boundary is an exfiltration route with a friendly interface.
Data leakage happens at the point of use. Staff may paste customer records, salary data and draft contracts into consumer tools without checking whether inputs are retained, used to improve models or processed outside the UK. The organisation may then have disclosed personal or confidential information, and potentially made an international transfer, without recording or assessing either.
Model drift is slower. A model that scored well at deployment degrades as the world it was trained on moves, and without monitoring, the first sign is a rise in complaints or losses several quarters later.
Third-party exposure is where most organisations actually have an issue. You are buying software from suppliers who buy model access from someone else, so your contract needs training-lawfulness representations, notification of model updates, audit rights and indemnities. The information gap between you and the supplier is wide, and the contract is the only thing that closes it.
AI risk and cyber risk are different problems
A cyber risk assessment asks whether an attacker can get in. An AI risk assessment asks whether a system working exactly as designed can still produce an unlawful, unfair or damaging outcome. A recruitment model that has never been breached can still discriminate against every applicant it screens, and penetration testing will not find that, because nothing is broken in the sense a penetration test understands.
The controls differ too. Cyber controls are access, patching, segmentation, monitoring and response, whereas AI controls are bias testing, human oversight design, model documentation, retraining triggers, output logging, red teaming for prompt injection and a route for a person to contest a decision. The two overlap on data protection and vendor management, so a mature security programme is a head start rather than a substitute.
What an AI risk assessment produces
The deliverables have to be usable by the board, by a regulator and by the engineer who has to fix something. A complete assessment gives you:
- An AI inventory covering every system in use, including features embedded in software you already pay for, with the business owner, the data it processes and whether you are provider or deployer.
- A classification of each system by risk, mapped to categories a regulator recognises rather than an internal scale nobody else understands.
- A risk register with likelihood, impact, current controls and residual risk, maintained rather than produced once.
- A gap analysis against a named standard, so a finding reads as missing human oversight documentation rather than as a vague sense that a system feels risky.
- A mitigation plan with owners, dates and a cost estimate per item.
- A monitoring and review regime: who checks performance, how often, and what triggers a reassessment.
The inventory is what people underestimate. It is tedious, it takes longer than expected, and everything else depends on it. An assessment that skips to policy writing produces a document describing a business you do not have.
The standards to focus on
Three reference points do most of the work, and they answer different questions.
ISO/IEC 42001, published in December 2023, is the AI management system standard. It is certifiable, which matters commercially: a certificate is something you can put in front of an enterprise customer, an insurer or a procurement team. The structure is familiar to anyone who has run an ISO 27001 programme.
ISO/IEC 23894 is guidance on AI risk management and maps onto ISO 31000, so it fits organisations already running enterprise risk in that shape. It is not certifiable and works best as the method behind the assessment.
The NIST AI Risk Management Framework, published in January 2023 with a Generative AI Profile added in July 2024, is voluntary, free and organised around four functions: govern, map, measure and manage. It is a practical starting point for a business that wants structure before committing to certification.
Why AI governance matters in London
London’s economy concentrates in exactly the sectors the EU AI Act treats as high risk and the UK regulators watch most closely.
Financial services and insurance use AI for credit scoring, fraud detection, underwriting and pricing, but their treatment under the EU AI Act depends on the particular use. AI used to assess the creditworthiness or credit score of natural persons can fall within Annex III, as can risk assessment and pricing for natural persons in life and health insurance, while systems used solely to detect financial fraud are expressly excluded from the creditworthiness category.
Banks may also face existing PRA model risk expectations. Recruitment and HR technology sits within an employment category scrutinised by the ICO. Legal and professional services firms face confidentiality and privilege questions when documents are submitted to third-party models, alongside potential professional negligence exposure. Healthcare organisations frequently process special category data, while edtech providers may handle children’s personal data and support decisions affecting them.
If you sell into the EU from London, the strategic question is whether to run UK and EU regimes in parallel or build to the EU position once and carry the cost across the whole product. That is commercial as much as compliance, and it is easier to answer with an inventory in front of you.
What it costs and how long it takes
Cost is driven by the number of systems in scope, whether personal data is involved, whether you are provider or deployer, whether you have EU exposure, and which sector regulator has an interest in you. Six AI tools with no EU customers and no automated decisions about individuals is a different engagement from a lender with models in the credit path.
For most mid-sized organisations the discovery and inventory phase runs in weeks rather than months, and the work that follows depends entirely on what discovery finds. Treat a fixed price quoted before anyone has seen your inventory as a guess. Ask for a fixed price on discovery and a rate card for what comes after. Budget for it as a cycle rather than a one-off, because a new system or a material model update should trigger a reassessment.
How to choose an AI governance consultant
The market has filled with AI governance offerings, many from firms that added the phrase to an existing service page. Six questions separate the useful from the decorative.
- Which standard will you assess against, and can I see a redacted example of the output?
- How do you discover shadow AI, and what happens if we find systems the executive team did not know about?
- Who on your team has run a data protection impact assessment on an AI system, and can I speak to them directly?
- How do you handle the split between provider and deployer obligations under the EU AI Act for our specific systems?
- What do you hand over so we can maintain the risk register ourselves, and what does ongoing support cost?
- Which professional bodies, certification schemes or regulatory regimes apply to your organisation and the individuals carrying out the work?
That last question is fair to ask of anyone about to see your systems. Cybercy Group is Cyber Essentials Plus certified, registered with the ICO under ZB014816, and an advisory partner to the Cyber Resilience Centre, working across cybersecurity, data protection and AI governance in the UK, Europe, the UAE and India.
Frequently asked questions
How often should AI risk assessments be performed?
Annually as a baseline, with an interim reassessment whenever a new system is introduced, a model is materially updated, a system starts being used for a purpose it was not assessed for, or the regulatory position changes. The appropriate frequency will depend on how quickly the organisation’s systems, use cases and regulatory exposure change.
Who should conduct an AI risk assessment?
Someone independent of the team that built or bought the system, which internally means risk, compliance or data protection rather than the product owner. Many organisations use an external assessor for the first cycle, partly for the method and partly because an outsider finds it easier to ask why a tool went live without review.
Are AI risk assessments mandatory in the UK?
There is no standalone UK requirement under that name. If your AI processes personal data and is likely to result in high risk to individuals, UK GDPR requires a data protection impact assessment covering much of the same ground. EU exposure brings documented risk management for high-risk systems under the AI Act, and regulated financial services firms already carry model risk expectations.
Is using ChatGPT a business risk?
Using it is not the risk. Using it without knowing who is using it, on what data and under which terms is. The workable answer for most organisations is an approved tool on a business agreement, a clear rule on what may be entered into it, and light monitoring, rather than a ban that pushes usage onto personal accounts where you have no visibility at all.
What documents come out of an assessment?
An AI inventory, a classification and risk register, a gap analysis against your chosen standard, a mitigation plan, a staff AI use policy and a board summary. High-risk systems also need technical documentation, human oversight procedures and records of testing. The precise documentation required depends on the organisation’s role and the system involved.
Where to start this week
Book an hour and list every tool your business uses that has AI in it, including the ones you did not buy for AI: the CRM that now scores leads, the Microsoft tenancy with Copilot enabled, the note-taker in your calendar invitations, the applicant tracking system with ranking switched on. Then mark which touch personal data and which sit in a decision that affects a person.
Most people are surprised by the length of that list, and the surprise is the point. It is the first page of the assessment, and the page that tells you how much of the rest you need.
Cybercy Group’s AI Governance and Enablement service covers AI strategy, use case identification, ethical AI framework development and performance monitoring. For a faster read on where you stand, the Cybercy+ Check gives you a scored assessment and a prioritised report.
